This paper reveals a characteristic of DEtection Transformer (DETR) that negatively impacts its training efficacy, i.e., the cross-attention and self-attention layers in DETR decoder have opposing impacts on the object queries (though both impacts are important). Specifically, we observe the cross-attention tends to gather multiple queries around the same object, while the self-attention disperses these queries far away. To improve the training efficacy, we propose a Divide-And-Conquer DETR (DAC-DETR) that separates out the cross-attention to avoid these competing objectives. During training, DAC-DETR employs an auxiliary decoder that focuses on learning the cross-attention layers. The auxiliary decoder, while sharing all the other parameters, has NO self-attention layers and employs one-to-many label assignment to improve the gathering effect. Experiments show that DAC-DETR brings remarkable improvement over popular DETRs. For example, under the 12 epochs training scheme on MS-COCO, DAC-DETR improves Deformable DETR (ResNet50) by +3.4AP and achieves 50.9 (ResNet-50) / 58.1 AP (Swin-Large) based on some popular methods (i.e., DINO and an IoU-related loss).

Self-attention architectures, which are rapidly pushing the frontier innatural language processing, demonstrate asurprising depth-inefficient behavior: previous works indicate that increasing the internal representation (network width) isjust as useful as increasing the number of self-attention layers (network depth).

However, these models suffer from quadratic computational cost in the input sequence lengthn to compute pairwise attention in each layer.

Transformer architectures have shown impressive performance in multiple research domains and have become the backbone of many neural network models. However, there is limited understanding on how Transformer works. In particular, with a simple predictive loss, how the representation emerges from the gradient training dynamics remains a mystery. In this paper, we analyze the SGD training dynamics for 1-layer transformer with one self-attention plus one decoder layer, for the task of next token prediction in a mathematically rigorous manner. We open the black box of the dynamic process of how the self-attention layer combines input tokens, and reveal the nature of underlying inductive bias. More specifically, with the assumption (a) no positional encoding, (b) long input sequence, and (c) the decoder layer learns faster than the self-attention layer, we prove that self-attention acts as a discriminative scanning algorithm: starting from uniform attention, it gradually attends more to key tokens that are distinct for a specific next token to be predicted, and pays less attention to common key tokens that occur across different next tokens. Among distinct tokens, it progressively drops attention weights, following the order of low to high co-occurrence between the key and the query token in the training set. Interestingly, this procedure does not lead to winner-takes-all, but decelerates due to a phase transition that is controllable by the learning rates of the two layers, leaving (almost) fixed token combination. We verify this scan and snap dynamics on synthetic and real-world data (WikiText).